Seismic reservoir prediction plays an important role in oil exploration and development. With the progress of artificial intelligence, many achievements have been made in machine learning seismic reservoir prediction. However, due to the factors such as economic cost, exploration maturity, and technical limitations, it is often difficult to obtain a large number of training samples for machine learning. In this case, the prediction accuracy cannot meet the requirements. To overcome this shortcoming, we develop a new machine learning reservoir prediction method based on virtual sample generation. In this method, the virtual samples, which are generated in a high-dimensional hypersphere space, are more consistent with the original data characteristics. Furthermore, at the stage of model building after virtual sample generation, virtual samples screening and model iterative optimization are used to eliminate noise samples and ensure the rationality of virtual samples. The proposed method has been applied to standard function data and real seismic data. The results show that this method can improve the prediction accuracy of machine learning significantly.